The manufacturing sector has been influenced by Industry 4.0. With the growing need for flexible, customized and on-demand services, the Manufacturing-as-a-Service (MaaS) paradigm has gained increasing relevance. This paper presents a comprehensive literature review of the use of AI in MaaS environments. The content analysis is structured around six focal areas: (i) dominant research methodologies; (ii) the spectrum of industries adopting MaaS; (iii) core enabling technologies, such as the internet of things, cloud computing, digital twins or blockchain; (iv) the classification of AI algorithms, including machine learning; (v) evolving objectives in MaaS and AI integration, and (vi) the interplay between MaaS and collaborative networks. The findings reveal the critical role of AI in MaaS ecosystems, enabling intelligent and service-oriented manufacturing.

The emergence of Industry 4.0 and Industry 5.0 dictates the need for intuitive, flexible, and intelligent human-machine interaction systems. This paper presents a modular approach to enhancing Digital Intelligent Assistants (DIAs) in manufacturing through the integration of data-driven and Large Language Model (LLM)-based skills. In the context of voice assistants, a "skill" refers to a specific functionality or capability that the assistant can perform. Specifically, we propose and implement three modular skills: (i) a “production reporting” skill, which enables voice-driven querying of ERP data for insights into inventory and planning; (ii) an “asset health monitoring” skill, which analyzes asset sensor data to monitor equipment condition and critical events; (iii) a “quality control” skill, which ingests unstructured documentation into a structured relational database and answers natural-language questions through LLM-based SQL translation. All the aforementioned three skills are integrated into a modular DIA framework that supports secure, multi-device interaction through natural language. The framework is applied to three real-life manufacturing scenarios, demonstrating how voice-enabled interfaces can improve decision-making and operational efficiency across diverse manufacturing workflows.

In the era of the digital economy, artificial intelligence technology has been deeply embedded in all aspects of manufacturing systems. AI-industrial technology convergence provides an important enabling pathway for enterprises to achieve breakthrough innovation. Taking Chinese listed manufacturing companies as the research sample, this study empirically examines the impact of AI-industrial technology convergence on firms’ breakthrough innovation and its boundary conditions. The study finds that AI-industrial technology convergence significantly improves the level of firms’ breakthrough innovation. Financial flexibility and the degree of technological blockade positively moderate the relationship between AI-industrial technology convergence and breakthrough innovation. Heterogeneity analysis shows that in high-tech industries, the promoting effect of AI-industrial technology convergence on breakthrough innovation is more significant than in non-high-tech industries. The conclusions of this study expand research on the innovation effects of artificial intelligence and technology convergence, providing theoretical foundations and practical guidance for enterprises to deepen the convergence of digital and industrial technologies and achieve innovation leaps, and also offering policy references for governments in formulating innovation support policies.

As machine learning (ML) systems increasingly govern high-stakes manufacturing processes, the demand for Explainable AI (XAI) has surged. While local, model-agnostic methods like LIME and SHAP are widely adopted to interpret opaque algorithms, they often suffer from severe instability and sensitivity to the inherent noise of industrial data. This paper addresses the "reliability gap" in operational XAI by proposing and evaluating two novel interpretability enhancements. First, we introduce CESHAP, a complexity-aware approach that integrates feature interactions to stabilize global feature importance rankings in non-linear models. Second, we propose AKDE-LIME, a novel local surrogate model utilizing Adaptive Kernel Density Estimation to weight perturbations based on the local data manifold, significantly reducing sampling variance. To quantitatively assess these methods, we utilize NAFIC (Normalized Average Feature Importance Concordance), an evaluation metric for noise robustness. Extensive comparative analysis across seven ML architectures (including Random Forest, XGBoost, and LightGBM) demonstrates that AKDE-LIME and CESHAP achieve stability and robustness scores orders of magnitude superior to standard baselines. These advancements provide the reliability necessary for deploying trustworthy AI in high-stakes industrial environments.

This paper aims to contribute towards the automation of information policies evaluation against the ISO/IEC 27001:2022 provisions. Thus, the study introduces an evaluation of different subsets of state-of-the-art natural language processing models (NLP) for compliance detection within information security policies. The analysis was carried out in form of an exploratory experiment, which entailed comparing zero-shot performance of pre-trained inference models, semantic similarity models and generative models across three compliance tasks. Results indicate that generative models perform better in requirement detection and contradiction identification, while certain NLI models offer more reliable contradiction detection at lower computational costs. No single model generated consistent results across the evaluated tasks. This confirms the need for adopting a multi-architecture approach as well as highlights the need for a domain-specific fine-tuning. This research is a work in progress for the development of an information security policy automatic compliance checker.